(235f) Predicting the 3D Printability of Polymer-Based Ink Using Machine Learning.

3D printing enhances the utilization of materials more efficiently than conventional manufacturing methods and facilitates the creation of intricate designs customized to meet specific requirements. Additionally, it enables the quick production of prototypes, expediting the development of new products. Nevertheless, the creation of 3D-printed items necessitates much experimentation and refinement. The lack of knowledge regarding the properties of the polymer material used in ink production and its subsequent influence on the result of 3D printing is the cause of this trial-and-error effort. As a result, we need to change the process parameters for successful printing if we change the material in the ink preparation. This work aims to establish a machine learning model framework capable of accurately predicting the outcome of extrusion-based 3D printing using polymer-based ink for high-value product (i.e., pharmaceuticals) applications. The goal is to reduce the need for trial and error, eliminate material waste, and ultimately lower costs. As a demonstration of feasibility, we formulated 24 distinct ink samples using a range of material compositions. We then conducted 220 observations to analyze the effects of changes in the process parameters. We conducted three fundamental rheology tests on each ink and analyzed their properties in terms of pre-printing stage, in-process condition, and post-printing stage. The rheology test findings are converted into data and used as input for the model. Next, we input the data into machine learning (ML) models (Random Forest and Artificial Neural Network) incorporating process variables and assess the performance of each ML model in terms of predictability. The results of our study demonstrate significant progress in 3D printing prediction, with accuracy rates of 71.4% and 63.3%, respectively. These findings indicate a high level of accuracy and successful prediction. The F1 score for Random Forest and Artificial Neural Network is 79% and 76%, respectively, validating our proof of concept using limited experimental data.